3D Point Cloud Ground Feature Detection for HD Map Updates
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional maps lack the precision and accuracy required for safe navigation of autonomous vehicles, and existing mapping techniques are expensive, time-consuming, and unable to provide up-to-date data, posing challenges for autonomous vehicle navigation.
Innovation Solution
The use of high-definition (HD) maps generated using deep learning-based models to detect ground features from sensor data, allowing autonomous vehicles to navigate with high accuracy and up-to-date information, leveraging data from lower resolution sensors and a distributed computing system for efficient map management.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional maps are used, then mapping cost and time are reduced, but map accuracy and precision fall below the safety threshold
Solution Approach 1:
The patent uses sensor data from vehicles (cameras, LIDAR, radar) to create a digital copy of the physical environment, replacing traditional survey methods. This digital twin approach enables high-precision mapping without requiring expensive survey equipment while maintaining accuracy within the safety threshold.
Solution Approach 2:
The patent replaces mechanical survey teams with automated sensor-based systems. Instead of human surveyors using physical measurement tools, the system uses electronic sensors mounted on vehicles to automatically capture and process environmental data, significantly reducing manual labor and time requirements while improving map freshness.
2Reliability
If traditional survey teams create maps, then initial map creation is possible, but map updates cannot keep pace with road changes
Solution Approach 1:
The patent implements continuous map updating by having vehicles continuously capture and transmit sensor data as they drive through areas. This ongoing data collection ensures the map is constantly refreshed with the latest road conditions, eliminating the periodic updates inherent in traditional survey methods and ensuring map freshness matches real-world changes.
Solution Approach 2:
The system uses feedback from sensor data to automatically detect and report road changes. When sensors detect modifications to roads, intersections, or other infrastructure, the system processes this feedback information and updates the map accordingly, creating a responsive mapping system that adapts to real-time changes without manual intervention.
3Measurement precision
If high resolution sensors are deployed, then map accuracy improves, but system cost and data storage requirements increase
Solution Approach 1:
The patent segments the mapping task by using multiple lower-resolution sensors (cameras, LIDAR, radar) distributed across multiple vehicles rather than relying on a single high-resolution system. Each sensor captures specific information (visual, spatial, radar data), and the system integrates these segmented data sources to achieve the required location accuracy while managing data volume through targeted collection.
Solution Approach 2:
The system collects more data than strictly necessary (excessive action) from multiple sensors, then processes and filters this data to extract only the essential information needed for navigation. This approach ensures comprehensive coverage and redundancy for safety while avoiding the need for expensive high-resolution sensors, as the necessary accuracy is achieved through intelligent data processing rather than raw data volume.
Data Source
AI summary
According to an aspect of an embodiment, operations may comprise accessing a map of a region. The map may comprise a three-dimensional (3D) point cloud comprising points of the region. Each of the points may comprise a pixel that comprises a red channel, a green channel, a blue channel, and intensity channel, and a height channel. The operations may also comprise identifying a set of points on the ground in the 3D point cloud. The operations may also comprise providing the set of points as input to a deep learning model. The operations may also comprise determining, using the deep learning model, whether each of the set of points is within a ground feature. The operations may also comprise storing, in the map, information regarding the ground feature in the map of the region.